These two get used interchangeably in agent-memory discussions, and they are not the same idea. They are not even the same kind of idea.
An ordinary graph makes two commitments at once. Every edge joins exactly two vertices, and the whole thing lives in ordinary Euclidean space. Loosen either commitment and you get a different structure. Loosen the other and you get another one. They are orthogonal, so there are four boxes, not two.
| Structure | An edge joins | Lives in | What it buys |
|---|---|---|---|
| Plain graph | 2 vertices | Euclidean | general pairwise links |
| Hypergraph | many vertices | Euclidean | n-ary facts as one relation |
| Hyperbolic graph | 2 vertices | hyperbolic | hierarchy with low distortion |
| Hyperbolic hypergraph | many vertices | hyperbolic | both at once |
Two different pressures push you to two different rows, and knowing which pressure you are under is most of the decision.
Hierarchy pressure pushes you to change the space
Euclidean space embeds deep hierarchies badly, and the reason is dimensional rather than incidental. A tree's node count explodes exponentially with depth while flat volume grows only polynomially, so nodes pile up on each other and the distances start lying to you.
Hyperbolic space fixes that by construction: volume grows exponentially with radius, which is the shape a tree already has.
This part is settled prior art and none of it is new. A weighted tree embeds in the hyperbolic plane with arbitrarily small distortion (Sarkar, 2011), that embedding reaches near-perfect precision in as few as two dimensions (Sala et al., 2018), and learned Poincaré embeddings beat Euclidean ones on hierarchical data (Nickel and Kiela, 2017). Hyperbolic graph convolutions cut link-prediction error by up to 63.1 percent in ROC AUC against Euclidean GCN baselines (Chami et al., 2019, arXiv:1910.12933).
A hyperbolic graph puts that to work while leaving edges alone: still two vertices per edge, laid out in the Poincaré ball instead of flat space.
So the first cue is short. If the dominant relation in your memory is "X is part of, or a kind of, Y", you want the hyperbolic axis, and you change the space rather than the edge.
The wall in front of that axis is not mathematical
Here is the part that does not show up in the papers, and it is the reason this axis is mostly theoretical in production today.
Every mainstream production text-embedding model emits Euclidean vectors compared by cosine similarity. OpenAI text-embedding-3, Cohere Embed, Voyage, Gemini Embedding, the open sentence-transformer family. I could not find a widely available model that natively outputs Poincaré-ball vectors.
Native hyperbolic NLP exists and is moving, but it is research-stage: Poincaré GloVe, hyperbolic fine-tuning (HypLoRA), and the first billion-parameter fully-hyperbolic language models (HELM, NeurIPS 2025).
And the geometry does not bolt on afterwards. The obvious workaround is to read a Euclidean vector as a tangent vector at the origin and apply the exponential map into the ball. The HypLoRA analysis shows what that actually does: it chains maps that cancel, and the composition reduces to an approximately Euclidean transformation. An encoder trained in Euclidean space never learned hyperbolic structure, and projecting its output does not teach it any.
That is a real constraint, not a preference. Anyone claiming hyperbolic memory on top of a standard embedding model is claiming a geometry their encoder does not have.
N-ary pressure pushes you to change the edge
The other axis leaves space flat and changes what an edge is. A hyperedge is a set: one relation joining many vertices at once.
As the HyperGraphRAG authors put it, "each edge in an ordinary graph connects only two entities, limiting their ability to represent the n-ary relations" (Luo et al., NeurIPS 2025, arXiv:2503.21322).
For agent memory this is the who-did-what-with-what-and-got-what problem. "Agent A used tool T on resource R in session X and got outcome O" is one fact about five entities, and a pairwise graph has two ways to stand in for it.
Clique expansion wires all five participants to each other. It is lossy in a specific and nasty way: a five-clique is indistinguishable from five unrelated pairs, so membership in the relation is simply gone.
Star expansion adds one node standing in for the relation and links every participant to it. This one is lossless for membership, because the encoding is a bijection and you can always recover which entities belong to which relation.
So the second cue: if your facts routinely bind three or more entities into one irreducible relation, you are feeling n-ary pressure, and you change the edge rather than the space.
The bridge most people miss
You do not need a hypergraph to store an n-ary fact on a pairwise graph.
Star expansion has been reinvented under several names because it keeps being the right answer: it is the incidence or Levi graph in graph theory, the W3C n-ary relation node, RDF reification, and the event node of formal semantics.
The price is concrete and small. One extra node per relation. One hop of indirection, so participants sit at distance two rather than adjacent. And a plain graph model will not exploit hyperedge-level structure the way a purpose-built hypergraph network would.
Which means n-ary is not a capability wall for a pairwise engine. It is a modelling choice with a known, cheap, lossless bridge if the workload ever demands it. The same cannot be said for the hyperbolic axis, where the bridge provably collapses.
That asymmetry is the practical takeaway. One of these axes you can defer and cross later. The other one you cannot fake.
Both at once
Because the axes are orthogonal, you can loosen both. A hyperbolic hypergraph has many-vertex hyperedges and curved space, which is the natural home for knowledge that is n-ary and hierarchical at the same time: a five-way event that also sits at a particular depth in a taxonomy.
It is an active research line rather than a shipping option. Hyperbolic hypergraph networks operate in hyperbolic space "to capture the tree-like hierarchy" while keeping the hyperedge's n-ary structure (Li et al., 2024, arXiv:2412.12158).
The two questions
Before reaching for either word:
- Is my dominant relation hierarchical? Then the pressure is on the space, and the honest state of the art is that native hyperbolic embeddings are not yet available off the shelf.
- Do my facts bind three or more entities irreducibly? Then the pressure is on the edge, and star expansion stores them losslessly on a pairwise graph today.
Answering "both" is legitimate, and it puts you in a research area rather than a product decision. Answering neither means a plain graph is doing its job and the vocabulary is not your problem.
Disclosure: I work on Mnemoverse, a memory engine for AI agents. Every figure above is from a linked paper by other people, and the 63.1 percent is Chami et al. on their own benchmarks, not a measurement of any product. The longer version is on the original page.
Top comments (0)